SearchEyes Boosts Multimodal Deep Search Intelligence with Simulation

Training multimodal search agents to perform complex, multi-hop reasoning has been difficult because existing development pipelines often treat training data, search environments, and reward signals as separate components. This leads to discarded structural metadata, reliance on external search engines, and sparse rewards for reinforcement learning. To overcome these issues, SearchEyes proposes a novel approach using a "simulated search world" grounded in a typed knowledge graph. This unified framework integrates all three components. It introduces Perception-Knowledge Chains (PKC) to sample constrained multi-hop paths over the visual-knowledge intersection of Wikidata5M, preserving crucial hop-level entity metadata. This metadata simultaneously defines a self-contained search world and provides precise step-level reward anchors. Furthermore, SearchEyes incorporates Hop-Anchored Policy Optimization (HaPO), which directly reuses these anchors for step-level credit assignment, eliminating the need for a separate reward model. Experiments across six multimodal knowledge-intensive benchmarks demonstrated that SearchEyes achieves state-of-the-art performance among open-source multimodal search agents, with its 27B model improving significantly over the strongest baseline.